🔍 Read the full analysis: I Fed Jev 30 Of My Images To See How It Responded on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
TypeSafe’s Jev 1.14 added opt-in image input and a bounding-box response type, with the company warning that the features are not ready for production use. Thorsten Meyer tested the model on 30 images from his site; product-image results were strong, but the supplied source cuts off before reporting the news-image findings or the full comparison with Jev 1.13.
TypeSafe added opt-in image input and a bounding-box response type to its Jev decision model on October 14, and a test published the next day found strong results on product photos. Thorsten Meyer tested 30 images from his site against headline-fit questions and product identification, while TypeSafe cautioned that the new capabilities remain too unreliable for production use without further evaluation.
Meyer tested the 30 most recent featured images on ThorstenMeyerAI.com: 10 product roundups, 15 news posts and five essays. He asked whether each image fit its own headline, a randomly selected different post headline and an unrelated headline about repotting a Monstera plant. The 10 product images also received a multiple-choice question asking which product each showed. Meyer said he made about 60 requests across Jev 1.14 and Jev 1.13.
For the product roundups, Jev 1.14 separated matching headlines from mismatches in the reported results. Images matched to their own headlines received scores from 0.65 to 0.89, with a median of 0.76. Random other headlines scored no higher than 0.18, with a median of 0.06, while the unrelated plant headline scored no higher than 0.10. Those figures describe this small test, not a general benchmark of image understanding.
On product identification, Jev 1.14 chose the correct item for nine of 10 images. Meyer described the miss as understandable: a photo showed an antenna-equipped router beside a cylindrical mesh unit, and the model selected “Wi-Fi 7 router” with confidence of 0.60. The correct answers each had confidence of at least 0.77. Jev 1.13 answered all 10 product questions correctly, though two answers had confidence below 0.5. The source text provided for this article ends as Meyer begins describing the news-image results, so those findings cannot be reported here.
Jev 1.14 · images + box
Jev can see now. Here is what 30 of my images showed.
I asked about every featured image three times: does it fit its own headline, a random other post’s headline, an off-topic headline? Both Jev versions, about 60 requests.
Does this image fit the headline?
Median “yes” probability on Jev 1.14, by kind of post
- Product roundups10 images0.760.061.13: 0.79 / 0.05
- News posts15 images0.330.161.13: 0.51 / 0.25
- My essays5 images0.510.161.13: 0.62 / 0.33
Off-topic headline (“How to Repot a Monstera Plant”): never above 0.10 on either version, 60 of 60 rejected.
Why news is weak: almost all news images show the same scene, people at desks looking at monitors. Jev could not tell them apart, and it was right not to.
1.14 is stricter than 1.13
Same 30 images, same questions
| Result | 1.14 | 1.13 |
|---|---|---|
| Fits own headline (≥ 0.5) | 15/30 | 23/30 |
| Fits a wrong headline (≥ 0.5) | 1/30 | 4/30 |
| Own beats random other (AUC) | 0.91 | 0.89 |
| Names the product (10 options) | 9/10 | 10/10 |
1.14’s one product miss: mesh system read as “Wi-Fi 7 router”, confidence 0.60. Re-tune thresholds before jev-latest moves.
What shipped on Oct 14
- headerImagesin state, instructions or criteria · 1.13 + 1.14 ·
TypeSafe-Danger-Zone-Images - 1.14 onlyBox typeone bounding box + confidence. 6 of 6 clear targets boxed (0.66–0.90); an absent bicycle still got a box (0.03)
- headerLong context64k / 32k → 262k tokens, slower and less accurate
- opt-inHelp find weaknessesqueries used for testing, not training
now: jev-preview = 1.14~Oct 21: jev-latest → 1.14notice: 60 days (1.13), 30 days (1.14)
Source: own test on October 15, 2026. 30 featured images from ThorstenMeyerAI.com (10 product roundups, 15 news posts, 5 essays), JPEG at up to 1,568 px, jev-1.14.0 and jev-1.13.0 with the image header. Bars show medians. AUC = probability that an image’s own headline scores higher than a random other headline. Features, prices and limits: TypeSafe blog post “Jev 1.14” and docs.typesafe.ai, as of October 15, 2026.
Powered by Thorsten Meyer AI
Image Checks Need Human Review
The test illustrates a possible use for image-capable decision models: repeatedly checking whether a website image fits its headline or depicts the intended product. Such checks could help publishers and online retailers review large image libraries, but Meyer’s results do not establish how well the approach works beyond his sample of 30 site images.
The confidence scores also show why a good-looking result should not be treated as a guarantee. In this test, the model’s incorrect product identification carried lower confidence than its correct answers. That pattern may help guide review in this particular task, but the source does not establish a reliable threshold for other images or uses. TypeSafe itself says the feature has enough unreliability to require an explicit opt-in and should not yet be trusted for production cases.
As an affiliate, we earn on qualifying purchases.
Jev Adds Images by Opt-In
Before the release, Jev accepted typed questions about text and returned structured answers such as yes-or-no probabilities, choices from a list or scores on a scale. In its October 14 release, TypeSafe said image support works with Jev 1.13 and 1.14, and described 1.13 as having been “multimodal all along.” The new bounding-box type, which returns one box for a described target with a confidence score, is limited to Jev 1.14.
Image processing is activated with the request header TypeSafe-Danger-Zone-Images: true. The company also put longer context behind a separate opt-in header, raising limits from 64,000 tokens per request and 32,000 per question to 262,000 for each. TypeSafe warns reliability falls on long inputs. The service accepts PNG and JPEG stills within stated size limits; the source says WebP and HEIC images must be converted first. These restrictions and warnings frame the image test as an early evaluation, not evidence that the features are production-ready.
“There is “still way too much jaggedness for us to vouch for these for production use cases.””
— TypeSafe, in its Jev 1.14 release announcement, as quoted by Thorsten Meyer
As an affiliate, we earn on qualifying purchases.
News Image Results Are Missing
The provided source ends mid-sentence as Meyer starts the section on news-image results. It does not include those scores, the essay-image findings, or a complete account of Jev 1.13’s headline-fit performance. No conclusions about those parts of the test can be drawn from the available text.
The reported results also come from a single publisher’s images and a small sample, with no independent evaluation described. The source does not give enough information to establish whether the scores generalize to other websites, image styles or subject areas. TypeSafe’s warning that the features are not ready for production remains relevant, and the test does not override it.
image bounding box annotation tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Model Rollout and Further Testing
TypeSafe said jev-preview was available at release and planned to move jev-latest to Jev 1.14 about a week later, around October 21, 2026. The company also gave notice periods for model deprecation: 60 days for Jev 1.13.0 and 30 days for Jev 1.14.0, with a formal policy still to come, according to Meyer’s account.
For readers considering image checks, the immediate next step is to evaluate the feature on their own data and keep human review in place. A fuller assessment of Meyer’s experiment would require the omitted results for news images and the remaining test details. TypeSafe has not, in the supplied material, announced a date for removing the danger-zone opt-in or a point at which it will endorse image processing for production.
As an affiliate, we earn on qualifying purchases.
Key Questions
What changed in Jev 1.14?
TypeSafe added opt-in image input and a new bounding-box answer type. Images can be used in the request state, question instructions or criteria; the bounding-box feature is limited to Jev 1.14.
How did Jev perform on product images?
In Meyer’s test, Jev 1.14 matched product images to their own headlines with scores of 0.65 to 0.89, and identified the pictured product correctly in nine of 10 cases. These results are specific to his small sample.
Was Jev 1.14 better than Jev 1.13?
The supplied source reports that Jev 1.14 got nine of 10 product identifications right, while Jev 1.13 got all 10 right but gave two answers confidence below 0.5. The source does not provide a complete comparison across all questions.
Is TypeSafe recommending Jev image support for production?
No. TypeSafe said the new capabilities still have too much unreliability for it to endorse them in production use cases. Image support requires an explicit opt-in header.
What is missing from the reported test?
The supplied article text cuts off as Meyer begins discussing the news-image results. It also does not provide the essay-image findings, so the complete outcome of the 30-image test is unavailable here.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
